Converting Float to Double Explicitly

Misc.
1

So unless I explicitly convert to Doubles I get errors when running MCMC:

def model(x, y):
    x = torch.tensor(x).to(dtype=torch.double)
    y = torch.tensor(y).to(dtype=torch.double)
    β = pyro.sample("β", dist.Normal(torch.zeros(x.shape[1],y.shape[1]),torch.ones(x.shape[1],y.shape[1]))).to(dtype=torch.double)
    μ = torch.matmul(x,β)
    θ = pyro.sample("θ", dist.HalfCauchy(torch.ones(x.shape[1]))).to(dtype=torch.double)
    L = pyro.sample("L", dist.LKJCorrCholesky(x.shape[1], torch.tensor(1.))).to(dtype=torch.double)
    L_Ω = torch.mm(torch.diag(θ.sqrt()), L).to(dtype=torch.double)
    with pyro.plate("data", x.shape[0]):
        obs = pyro.sample("obs", dist.MultivariateNormal(loc=μ, scale_tril=L_Ω), obs=y).to(dtype=torch.double)

Am I the only one who thinks this is really annoying/ugly? Is there some reason that Pyro can’t do this conversion under the covers?

2

pyro works with both floats and doubles as long as you’re consistent.

you can make your life easier by using torch api like torch.set_default_dtype(torch.float64)

1 Like
3

Ah! Thanks for the tip. Definitely makes life easier (and prettier).